picklescan & ModelScan semantic bypass (ShadowPickle-class): memo-overwrite desync
Benign security-research PoC for the huntr Model File Format program. The payload only writes a nonce marker file (echo / io.open) — no real harm, no network, no secrets (huntr §benign).
Vulnerability
The genops static walk (both scanners) records memo[key]=the lexically-preceding opcode's arg; a MEMOIZE(None) poisons the scanner's memo to None while the pickle VM memoizes the real 'os'/'system' strings. The STACK_GLOBAL back-walk then resolves the poisoned memo -> picklescan sees a BENIGN ('None','None') global (exit 0, walks clean, no error) while the C VM executes os.system. Distinct MECHANISM from a parse-abort (defeats reconstruction, not the parser). Maps to ShadowPickle, arXiv:2607.17503.
Measured evidence (2026-07-24)
Against picklescan 1.0.5 + ModelScan 0.8.8: both fail to flag; on load the benign marker fires. Re-verify: python -m nomos_redteam scanbypass.
Reproduce
import torch # or: import pickle
obj = torch.load("model.pt", weights_only=False) # fires the benign marker at load
Scan first to confirm the bypass:
picklescan -p model.pt # exit 0 / no dangerous global
modelscan -p model.pt # no issue reported